Power load prediction method and system based on hybrid deep learning
By using a hybrid deep learning approach, combining multidimensional features and multipath neural network models, the problems of insufficient feature capture and robustness in power load forecasting are solved, achieving high-precision and flexible deployment of power load forecasting and supporting the intelligent operation of the power grid.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- JIANGSU LINYANG ENERGY CO LTD
- Filing Date
- 2025-11-27
- Publication Date
- 2026-04-24
AI Technical Summary
Existing power load forecasting technologies struggle to capture multi-scale time-series features in a coordinated manner. Their feature systems are incomplete, their models lack robustness, and their deployment flexibility is limited, making them difficult to adapt to complex power grid environments.
A power load forecasting method based on hybrid deep learning is adopted. By constructing a neural network model with a multi-dimensional feature set and an encoder-decoder structure, and combining causal convolution, long short-term memory network and simplified state space model, features at different time scales are extracted. Global context information is integrated through a fully connected fusion layer and a Transformer encoder, and future load values are output by combining exogenous feature projection and location encoding.
It significantly improves the accuracy and adaptability of load forecasting, ensures the numerical stability of the training and inference process, achieves high prediction accuracy and flexible deployment of the model in a variety of complex scenarios, and supports intelligent dispatching of power systems.
Smart Images
Figure CN121920584A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the fields of artificial intelligence and power system technology, and in particular to a power load forecasting system and method based on a multi-path hybrid deep learning architecture. Background Technology
[0002] Electricity load forecasting is the cornerstone of the safe and economical operation of the power grid. Its forecasting accuracy directly affects the formulation of power generation plans, electricity market transactions, and the efficient consumption of renewable energy. With the improvement of the intelligence level of the power system, load forecasting technology has evolved from traditional statistical models to modern machine learning models.
[0003] Early methods primarily employed time series models such as Autoregressive Integral Moving Average (ARIMA), which performed well for linear and stationary series. However, electricity load is influenced by multiple factors, including weather, holidays, and social activities, exhibiting significant nonlinear and non-stationary characteristics, limiting the prediction accuracy of traditional statistical models. Subsequently, classic machine learning algorithms such as Support Vector Machines (SVM) and Random Forests were introduced. While they can capture some nonlinear relationships, they suffer from complex feature engineering and limited model expressive power when dealing with large-scale, high-dimensional time series data, making it difficult to uncover deep temporal dependency patterns in load data. In recent years, deep learning models, such as Recurrent Neural Networks (RNN), Long Short-Term Memory Networks (LSTM), and Convolutional Neural Networks (CNN), have become mainstream techniques for load forecasting due to their powerful feature learning capabilities. LSTM and its variants excel at handling medium- to long-term dependencies but have limitations in capturing local fine-grained patterns and ultra-long-range dependencies, and incur significant computational costs. CNNs effectively extract local features but are weak in modeling the overall temporal order of the sequence. The Transformer model demonstrates an advantage in handling long sequence dependencies due to its self-attention mechanism, but its computational complexity increases quadratically with sequence length, and it sometimes lacks sufficient perception of local details. In summary, a single deep learning model struggles to simultaneously and efficiently capture the complex features of different time scales in a load sequence, such as "short-term local fluctuations, medium-term periodic patterns, and long-term global dependencies."
[0004] In engineering deployment, existing methods face the following challenges: Incomplete feature systems: Many models rely on basic features (such as historical load and temperature) that are insufficient to comprehensively describe the driving factors of load changes, lacking effective fusion of multi-source heterogeneous data (such as humidity, rainfall, and refined calendar information) and the construction of high-order cross-features. Insufficient model robustness: Real-world data commonly contains missing values, noise, and even outliers, while most models' data preprocessing processes are not robust enough. Numerical instability during standardization (such as zero variance) can easily lead to training failure or prediction distortion. Limited deployment flexibility: Many high-precision models have high computational resource requirements, making them difficult to adapt to diverse deployment scenarios from cloud centers to edge computing nodes, limiting their application in situations with high real-time requirements.
[0005] Therefore, there is an urgent need in this field for a new load forecasting solution that can collaboratively capture multi-scale time-series features, has a complete feature engineering system, and ensures high robustness of the training and inference process from the algorithm level, in order to cope with the increasingly complex power grid operating environment. Summary of the Invention
[0006] The purpose of this invention is to overcome the above-mentioned defects of the prior art and provide a hybrid deep learning prediction system and method with high accuracy, strong robustness, and the ability to comprehensively capture the multi-timescale features of the load sequence.
[0007] The technical solution of this invention is:
[0008] This invention provides a power load forecasting method based on hybrid deep learning, comprising:
[0009] S1. Data preprocessing steps: acquire historical load data, meteorological data, and calendar data; perform missing value imputation and standardization; and construct a multidimensional feature set.
[0010] S2, Model training steps: Train a neural network model with an encoder-decoder structure based on the preprocessed data;
[0011] The encoder employs a multi-path hybrid architecture, comprising three parallel paths: a causal convolutional path (CNN) for extracting local temporal features, a long short-term memory network (LSTM) for capturing medium- to long-term periodic patterns, and a simplified state-space model (Mamba) for modeling long-range dependencies. The initial input of the decoder consists of the projection of exogenous features from future time periods, positional encoding, and a learnable query bias parameter.
[0012] S3. Load forecasting step: Input the current load data of the target transformer area into the trained neural network model to obtain the load forecast value for a specified future time period.
[0013] Furthermore, S1 includes:
[0014] S11. Missing values in historical load data are filled using the average of previous and subsequent time points;
[0015] S12. Perform standardization processing, identify and correct outliers in historical load data and meteorological data using box plot rules, and align all data according to specified time points to form a unified time series.
[0016] S13. Construct a multi-dimensional feature set, including basic features, time period features, load-derived features, and interaction features.
[0017] Furthermore, the basic features mentioned in S13 include historical load values, temperature, humidity, rainfall, and holiday factors. The time period features are represented by sine and cosine coding to indicate daily and weekly cycles. The load-derived features include the load lag term, moving average, and first-order difference value. The interaction feature is the product of the average temperature and the load lag term.
[0018] Furthermore, after performing data standardization in S12, a zero-value protection mechanism is used to compare the standardization result with a very small positive value. If it is less than the value, it is set to 1 to obtain the processed data.
[0019] Furthermore, the encoder employs a multi-path hybrid architecture, wherein;
[0020] Causal convolutional path CNN uses multiple layers of causal convolution with specified kernel size and dilation coefficient, followed by normalization and activation functions to extract local temporal features;
[0021] The Long Short-Term Memory (LSTM) network path uses a single-layer LSTM network with a specified hidden layer dimension, followed by linear projection and layer normalization to capture medium- and long-term cyclical patterns.
[0022] The simplified state-space model path Mamba uses a specified convolutional kernel size to model long-range dependencies through gating and convolution operations.
[0023] Furthermore, for the three parallel paths of the multi-path hybrid architecture, the outputs of each path are concatenated along the feature dimension and integrated through a fully connected fusion layer, specifically including:
[0024] The multidimensional feature set is input into the causal convolution path, the long short-term memory network path, and the simplified state space model path respectively. The outputs of the three paths are concatenated in the feature dimension to obtain the concatenated feature.
[0025] The spliced features are integrated through a fully connected fusion layer to obtain fused features;
[0026] The fused features are refined into global context through multiple Transformer encoder layers to obtain the encoder output, which is then used as the decoder input.
[0027] Furthermore, the initial input of the decoder consists of the sum of exogenous feature projections from future time periods, positional encodings, and a learnable query bias parameter, specifically including:
[0028] Acquire meteorological and calendar data for future periods, and obtain the exogenous feature projection of the load data through the projection layer;
[0029] The exogenous feature projection is added to the positional encoding to obtain the initial input, which is then combined with the learnable query bias parameters to form the initial input of the decoder.
[0030] The encoder output is interacted with by a multi-layer Transformer decoder, and the load prediction is output through a multi-layer fully connected layer.
[0031] Furthermore, during the training of the model:
[0032] The neural network model was trained using preprocessed data, employing the Huber loss function, AdamW optimizer, and learning rate scheduler, with settings for initial learning rate, weight decay, batch size, maximum number of training epochs, and early stopping mechanism.
[0033] If the validation set loss does not improve within a specified epoch, the learning rate is reduced using the learning rate scheduler.
[0034] Furthermore, in the load forecasting step of S3, the model is updated when the MAPE value exceeds the threshold or when the data accumulates to a preset time.
[0035] A power load forecasting system based on hybrid deep learning includes:
[0036] A data preprocessing module is configured to perform the data preprocessing steps;
[0037] A model training module, configured to store and execute the model training steps;
[0038] A load forecasting module, configured to perform the load forecasting steps.
[0039] The beneficial effects of this invention are:
[0040] This invention provides a power load forecasting method based on hybrid deep learning. It constructs a multi-dimensional feature set and designs a neural network model with an encoder-decoder structure for load forecasting. The encoder employs a three-path parallel architecture combining causal convolution, long short-term memory networks, and a simplified state-space model to extract features at different time scales. Global contextual information is then integrated through a fully connected fusion layer and a Transformer encoder layer to ensure prediction accuracy. The decoder combines exogenous feature projection and positional encoding to output future load values, significantly improving the accuracy and adaptability of load forecasting and providing reliable support for power system scheduling and resource optimization. Specific advantages are as follows:
[0041] 1. Multi-scale feature capture capability: Through an innovative multi-path hybrid encoder, the model collaboratively captures short-term local patterns, medium-term periodic regularities, and long-term global dependencies of the load, fundamentally improving the model's performance capability.
[0042] 2. Excellent robustness: Through the zero-variance protection mechanism of the standardization process, the numerical stability of the training and inference processes is ensured, enabling the model to adapt to noisy data in real industrial environments.
[0043] 3. Flexible deployment capability: The modular design allows the model to be flexibly trimmed or lightweighted according to the resource constraints of the actual deployment scenario, making it easy to implement on devices at all levels, including cloud, edge, and terminal.
[0044] 4. High prediction accuracy: The combination of a comprehensive feature system and a powerful model architecture enables it to maintain high prediction accuracy in a variety of complex scenarios, providing reliable support for smart grids.
[0045] Other features and advantages of the present invention will be described in detail in the following detailed description section. Attached Figure Description
[0046] The above and other objects, features and advantages of the present invention will become more apparent from the more detailed description of exemplary embodiments of the invention in conjunction with the accompanying drawings, wherein the same reference numerals generally represent the same components in the exemplary embodiments of the invention.
[0047] Figure 1 A flowchart of the power load forecasting method based on hybrid deep learning of the present invention is shown.
[0048] Figure 2 The architecture diagram of the neural network model in this invention is shown.
[0049] Figure 3 A diagram of the multipath hybrid architecture of the encoder in this invention is shown.
[0050] Figure 4The diagram shows a comparison of the effectiveness of the present invention in load forecasting in a certain region. Detailed Implementation
[0051] Preferred embodiments of the invention will now be described in more detail with reference to the accompanying drawings. While preferred embodiments of the invention are shown in the drawings, it should be understood that the invention can be implemented in various forms and should not be limited to the embodiments set forth herein.
[0052] This invention provides a power load forecasting method based on hybrid deep learning, comprising:
[0053] S1. Data preprocessing step: Obtain historical load data, meteorological API and calendar data, including load data at the 15-minute granularity, daily maximum temperature, minimum temperature, average temperature, average humidity, rainfall, and the corresponding day of the week, weekday marking and holiday classification.
[0054] The data was cleaned and aligned. Missing values in the load data were imputed using the mean of the preceding and following time points, and outliers were identified and corrected using box plots. All data were aligned to 15-minute time points to form a unified time series.
[0055] Construct a multidimensional feature set, which includes:
[0056] Basic characteristics: historical load values, temperature, humidity, rainfall, and holiday factors.
[0057] Time cycle characteristics: daily and weekly cycles represented by sine and cosine encoding.
[0058] Load-derived characteristics: lag term (Lag), moving average (MA), and first-order difference.
[0059] Interactive feature: Construct the product of average temperature and hysteresis load.
[0060] S2. Model Training Steps: Train a neural network model with an encoder-decoder structure using the processed data. The encoder uses a multi-path hybrid architecture with an input length of 7 days (672 15-minute data points). The three parallel paths include:
[0061] CNN path: Use two causal convolutional layers with kernel size of 7 and dilation coefficients of 1 and 2 respectively, followed by LayerNorm and GELU activation to specifically extract local temporal patterns.
[0062] LSTM path: Single-layer LSTM, hidden layer dimension 128, followed by linear projection and LayerNorm to capture mid-term cyclical patterns.
[0063] Mamba path: Employs a simplified state-space model with a convolution kernel size of 64, and efficiently models long-range dependencies through Sigmoid gating and convolution operations.
[0064] The outputs of the three paths are concatenated along the feature dimension and then fused through a fully connected layer. The fused features are then refined for global context through a two-layer Transformer encoder layer.
[0065] The decoder receives the exogenous feature projections and location codes for the next two days, while introducing a learnable query bias parameter. Combined with the learnable query bias parameter, the decoder interacts with the encoder through two layers of Transformer decoder memory, and finally outputs the predicted load value through three fully connected layers (dimensions 128, 64, and 1 respectively).
[0066] Loss function: The Huber loss function is used, and its formula is as follows:
[0067]
[0068] Here, δ is set to 1.0, and this function combines the stability of MSE with the robustness of MAE to outliers.
[0069] Optimization strategy: Use the AdamW optimizer with an initial learning rate of 5e-4 and weight decay of 1e-4. Employ the ReduceLROnPlateau learning rate scheduler, halving the learning rate if the validation set loss does not improve within 8 epochs. Training: Batch size is set to 64, maximum training epochs are 120, and early stopping is enabled (patience value 20). Before training, the normalizer is protected against zero variance by setting values in the scale less than 1e-8 to 1.
[0070] S3. Load forecasting step: Input the latest sequence data of the target area into the trained model to obtain the load forecast results for the specified future period. The Mean Absolute Percentage Error (MAPE) is used as the core evaluation indicator, and its calculation formula is as follows:
[0071]
[0072] S4: The system supports cloud-edge collaborative deployment, provides API interfaces, deploys lightweight models on edge computing nodes to achieve local inference, and can be integrated with existing power grid dispatching systems and power systems.
[0073] In practical applications:
[0074] (I) Data Reception and Database Construction
[0075] Data preparation: Collect 15-minute load data for the region over the past two years.
[0076] Data reception and storage: Continuously monitor and receive real-time data streams pushed from the main station. The data packets contain the unique identifier (SN) of the predicted object and its corresponding measured load value. After being formatted and normalized, the received data is stored in an orderly manner in the local time-series database, forming the basic load data set required for prediction.
[0077] Metadata Management: To effectively support model training and version management, the system has built a dedicated metadata database. This database records key metadata for each prediction object (SN), including its total data length in the local database and the actual data length used during the training of the current model instance. This mechanism provides crucial information for model update strategies, data backtracking, and analysis.
[0078] (II) Model Calculation and Application Process
[0079] Model Building and Training: The system periodically (e.g., every morning) iterates through all prediction objects in the metadata database. For objects whose total historical data volume meets a preset threshold (e.g., no less than 365 days) and whose current model version is outdated (i.e., the difference between the total data length and the model training data length exceeds a set value K), the model training process is automatically triggered. The complete historical data of this object will be used to optimize the prediction model parameters. After training is complete, the system serializes and saves the new model locally, updates the model version, and synchronously updates the model training data length of this object in the metadata database.
[0080] like Figure 2 , 3 The diagram shown illustrates a multi-path hybrid architecture for the neural network model and encoder.
[0081] The encoder input length (IN_DAYS) was set to 7 days (7*96 = 672 time points), and the decoder output length (FORECAST_DAYS) was set to 2 days (192 time points). The causal convolutional path used two convolutional layers with a kernel size of 7 and dilation coefficients of 1 and 2 respectively. The LSTM path's hidden layer dimension was set to 128. The Mamba path's convolutional kernel size was set to 64. The model was implemented in the PyTorch framework using the AdamW optimizer with an initial learning rate of 5e-4 and an early stopping strategy. Robustness handling: After data standardization (StandardScaler), the scale array was actively checked, setting any value less than 1e-8 to 1, and the var array was similarly processed to completely avoid the risk of division by zero.
[0082] Building a multi-dimensional feature set: To construct a more accurate prediction model, the system queries the geographic coordinates associated with the prediction object's serial number (SN) before training or inference. Based on these coordinates, the system automatically calls an open meteorological API to obtain forecast data for key meteorological elements in the region (such as maximum / minimum temperature, average humidity, and rainfall). This meteorological data, associated with the SN, is stored as important external features in the local database, enriching the model input.
[0083] Inference and Feedback: The system periodically monitors the local database to check for new measured load values for each forecast object (SN). Upon detecting new data, the inference calculation process of the loaded model is immediately triggered. The inference process combines the latest load sequence with corresponding meteorological characteristics to calculate the load forecast for a specified future period (e.g., the next two days). The calculation results are pushed back to the main station system in real time via a dedicated data interface, completing the closed loop of the forecast service. If no new data is found, the process enters a dormant state, waiting for the next query cycle.
[0084] Forecast and Results: In this embodiment, the system successfully achieved automated load forecasting for the region, such as... Figure 4 The figure shows the predicted curve and the actual value curve. Statistically, the average absolute percentage error (MAPE) of its output day-ahead load forecast curve is 2.87%. Furthermore, due to the parallel collaborative design of the dual processes, data reception and model inference do not block each other, ensuring that the end-to-end delay from data to decision is less than 15 seconds, significantly improving the response speed and intelligence level of power grid dispatch.
[0085] The various embodiments of the present invention have been described above. These descriptions are exemplary and not exhaustive, nor are they limited to the disclosed embodiments. Many modifications and variations will be apparent to those skilled in the art without departing from the scope and spirit of the described embodiments.
Claims
1. A power load forecasting method based on hybrid deep learning, characterized in that, include: S1. Data preprocessing steps: acquire historical load data, meteorological data, and calendar data; perform missing value imputation and standardization; and construct a multidimensional feature set. S2, Model training steps: Train a neural network model with an encoder-decoder structure based on the preprocessed data; The encoder employs a multi-path hybrid architecture, comprising three parallel paths: a causal convolutional path (CNN) for extracting local temporal features, a long short-term memory network (LSTM) for capturing medium- to long-term periodic patterns, and a simplified state-space model (Mamba) for modeling long-range dependencies. The initial input of the decoder consists of the projection of exogenous features from future time periods, positional encoding, and a learnable query bias parameter. S3. Load forecasting step: Input the current load data of the target transformer area into the trained neural network model to obtain the load forecast value for a specified future time period.
2. The power load forecasting method based on hybrid deep learning as described in claim 1, characterized in that... S1 includes: S11. Missing values in historical load data are filled using the average of previous and subsequent time points; S12. Perform standardization processing, identify and correct outliers in historical load data and meteorological data using box plot rules, and align all data according to specified time points to form a unified time series. S13. Construct a multi-dimensional feature set, including basic features, time period features, load-derived features, and interaction features.
3. The power load forecasting method based on hybrid deep learning as described in claim 2, characterized in that... In S13, the basic features include historical load values, temperature, humidity, rainfall, and holiday factors. The time period features are represented by sine and cosine coding to indicate daily and weekly cycles. The load-derived features include the load lag term, moving average, and first-order difference value. The interaction feature is the product of the average temperature and the load lag term.
4. The power load forecasting method based on hybrid deep learning as described in claim 1, characterized in that, After data standardization is performed in S12, a zero-value protection mechanism is used to compare the standardization result with a very small positive value. If it is less than the value, it is set to 1 to obtain the processed data.
5. The power load forecasting method based on hybrid deep learning as described in claim 1, characterized in that, The encoder adopts a multi-path hybrid architecture, wherein; Causal convolutional path CNN uses multiple layers of causal convolution with specified kernel size and dilation coefficient, followed by normalization and activation functions to extract local temporal features; The Long Short-Term Memory (LSTM) network path uses a single-layer LSTM network with a specified hidden layer dimension, followed by linear projection and layer normalization to capture medium- and long-term cyclical patterns. The simplified state-space model path Mamba uses a specified convolutional kernel size to model long-range dependencies through gating and convolution operations.
6. The power load forecasting method based on hybrid deep learning as described in claim 1, characterized in that, For the three parallel paths in the multi-path hybrid architecture, the outputs of each path are concatenated along the feature dimension and integrated through a fully connected fusion layer, specifically including: The multidimensional feature set is input into the causal convolution path, the long short-term memory network path, and the simplified state space model path respectively. The outputs of the three paths are concatenated in the feature dimension to obtain the concatenated feature. The spliced features are integrated through a fully connected fusion layer to obtain fused features; The fused features are refined into global context through multiple Transformer encoder layers to obtain the encoder output, which is then used as the decoder input.
7. The power load forecasting method based on hybrid deep learning as described in claim 1, characterized in that, The initial input to the decoder consists of the sum of exogenous feature projections from future time periods, positional encodings, and a learnable query bias parameter, specifically including: Acquire meteorological and calendar data for future periods, and obtain the exogenous feature projection of the load data through the projection layer; The exogenous feature projection is added to the positional encoding to obtain the initial input, which is then combined with the learnable query bias parameters to form the initial input of the decoder. The encoder output is interacted with by a multi-layer Transformer decoder, and the load prediction is output through a multi-layer fully connected layer.
8. The power load forecasting method based on hybrid deep learning as described in claim 1, characterized in that, During the training of the model: The neural network model was trained using preprocessed data, employing the Huber loss function, AdamW optimizer, and learning rate scheduler, with settings for initial learning rate, weight decay, batch size, maximum number of training epochs, and early stopping mechanism. If the validation set loss does not improve within a specified epoch, the learning rate is reduced using the learning rate scheduler.
9. The power load forecasting method based on hybrid deep learning as described in claim 1, characterized in that... In the load forecasting step of S3, the model is updated when the MAPE value exceeds the threshold or when the data accumulates to a preset time.
10. A power load forecasting system based on hybrid deep learning, characterized in that, include: A data preprocessing module is configured to perform the data preprocessing steps; A model training module, configured to store and execute the model training steps; A load forecasting module, configured to perform the load forecasting steps.